A multi-agent collaborative self-evolution method and system based on ontology semantic network

Abstract

The application provides a multi-agent collaborative self-evolution method and system based on an ontology semantic network, which comprises the following steps: using an index learning module to precipitate fixed index templates and dynamic index templates from quality inspection high-score queries and manage hot data caches; using a knowledge reverse updating module to receive feedback signals and update ONN model layers; and using a user portrait module to construct and maintain multi-dimensional user portraits and provide personalized contexts when clarifying intents. In the above technical solution, a self-evolution closed loop based on quality inspection feedback driving is realized, so that the system can continuously learn from the success and failure of each query and automatically improve query accuracy and response efficiency.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a multi-agent cooperative self-evolution method and system based on ontology semantic networks. Background Technology

[0002] In the actual operation of intelligent query systems based on large language models, the following technical challenges exist for continuous optimization:

[0003] 1. AI query systems lack the ability to self-improve from execution feedback. Existing solutions mostly adopt a "one-time inference" model—each query independently completes the entire process from natural language understanding to query code generation, without learning from past successful or failed queries. Every time the same frequently asked question is asked, the system needs to re-execute the entire inference and code generation process, wasting computational resources and increasing the probability of errors. See Wu et al., “An Empirical Study of In-context Learning in LLMs for Machine Translation”, ACL 2024.

[0004] 2. Knowledge updates rely on manual intervention and lack automated mechanisms. When business rules in the ontology semantic network change, synonym mappings are missing, or default conditions need adjustment, existing systems typically rely on manual modifications to the knowledge base by administrators. This approach is slow to respond, prone to omissions, and lacks systematic error classification and change management, failing to meet the knowledge governance needs of enterprise-level systems.

[0005] 3. Lack of personalized adaptation in understanding user intent. Users with different roles (such as research management personnel, academic affairs office leaders, and project finance personnel in universities) have significant differences in the object types, common conditions, and dimensions they focus on when querying. However, the existing system uses the same intent understanding strategy for all users, resulting in a large number of unnecessary clarification questions, which reduces user experience and efficiency.

[0006] 4. The problem of repetitive calculations in high-frequency queries. In actual operation, a large number of queries have high repetition or structural similarity (such as monthly "teachers' paper statistics" and "student grade rankings"). However, the existing system performs complete ABC parsing and DSL generation from scratch every time, and lacks a mechanism to cache the validated query logic as a reusable template.

[0007] 5. Lack of a hot data caching strategy. Even if a caching mechanism exists, it usually caches the "data results" rather than the "query logic template". Data result caching faces real-time issues (the cache becomes invalid after the underlying data changes), while query logic template caching can re-execute the DSL to obtain the latest results after the underlying data changes, thus balancing caching efficiency and data real-time performance. Summary of the Invention

[0008] This application provides a multi-agent collaborative self-evolution method and system based on ontology semantic network to realize a self-evolution closed loop driven by quality inspection feedback, enabling the system to continuously learn from the success and failure of each query, and automatically improve query accuracy and response efficiency.

[0009] Firstly, a multi-agent cooperative self-evolution method based on ontology semantic networks is provided, including the following steps:

[0010] The indicator learning module is used to extract fixed and dynamic indicator templates from high-scoring quality inspection queries and manage hot data caching.

[0011] The knowledge-backward update module receives feedback signals and updates the ONN model layer.

[0012] Utilize the user profile module to build and maintain multi-dimensional user profiles and provide personalized context when clarifying intent.

[0013] In the above technical solution, a fixed indicator template and a dynamic indicator template are extracted from high-scoring quality inspection queries by using an indicator learning module and hot data cache is managed; a knowledge reverse update module is used to receive feedback signals and update the ONN model layer; a user profile module is used to build and maintain multi-dimensional user profiles and provide personalized context when intent is clarified; a self-evolutionary closed loop driven by quality inspection feedback is realized, which enables the system to continuously learn from the success and failure of each query and automatically improve query accuracy and response efficiency.

[0014] In one specific implementation scheme, it also includes:

[0015] The scheduling module is used to initiate fixed matching, dynamic matching, and deep parsing in parallel, and the optimal result is selected based on the coverage score.

[0016] In one specific implementation scheme, the indicator template includes:

[0017] Fixed indicator templates allow for direct reuse of complete OQS and DSL;

[0018] Dynamic indicator templates are reused through parameterized placeholders.

[0019] In a specific feasible implementation, knowledge reverse updates modify the ONN model layer in accordance with a safe change process that includes change logging, impact analysis, gray-scale verification, and rollback capability.

[0020] In a specific feasible implementation, the content of knowledge reverse update includes: supplementing synonym mapping, correcting CBC constraints, adding default rules, and improving business knowledge descriptions.

[0021] In a specific feasible implementation, user personas include dimensions such as the range of ontology categories commonly used by the user, the user's organizational and job descriptions, and the user's preferences for requesting the system to remember their past conversations.

[0022] In one feasible implementation, the object of hot data caching is the query logic template.

[0023] In a specific implementation plan, the high-scoring signal triggering indicator learning agent accumulates the template, while the low-scoring signal triggering knowledge learning agent updates the ONN in reverse.

[0024] Secondly, a multi-agent cooperative self-evolutionary system based on ontology semantic networks is provided, including:

[0025] The indicator learning module is used to extract fixed indicator templates and dynamic indicator templates from high-scoring quality inspection queries and manage hot data cache;

[0026] The knowledge reverse update module is used to receive feedback signals and update the ONN model layer;

[0027] The user profile module is used to build and maintain multi-dimensional user profiles and provide personalized context when intent is clarified.

[0028] In the above technical solution, a fixed indicator template and a dynamic indicator template are extracted from high-scoring quality inspection queries by using an indicator learning module and hot data cache is managed; a knowledge reverse update module is used to receive feedback signals and update the ONN model layer; a user profile module is used to build and maintain multi-dimensional user profiles and provide personalized context when intent is clarified; a self-evolutionary closed loop driven by quality inspection feedback is realized, which enables the system to continuously learn from the success and failure of each query and automatically improve query accuracy and response efficiency.

[0029] In one specific implementation scheme, it also includes:

[0030] The scheduling module is used to initiate fixed matching, dynamic matching, and deep parsing in parallel and select the optimal result based on the coverage score. Attached Figure Description

[0031] Figure 1A flowchart illustrating the multi-agent cooperative self-evolution method based on ontology semantic network provided in this application embodiment;

[0032] Figure 2 This is a structural block diagram of a multi-agent cooperative self-evolutionary system based on ontology semantic network provided in an embodiment of this application. Detailed Implementation

[0033] The present application will now be described in further detail with reference to the accompanying drawings and embodiments. Through these descriptions, the features and advantages of the present application will become clearer and more apparent.

[0034] The term “exemplary” as used herein means “serving as an example, embodiment, or illustration.” Any embodiment illustrated herein as “exemplary” is not necessarily to be construed as superior to or better than other embodiments. Although various aspects of embodiments are shown in the accompanying drawings, the drawings are not necessarily drawn to scale unless specifically indicated otherwise.

[0035] Furthermore, the technical features involved in the different embodiments of this application described below can be combined with each other as long as they do not conflict with each other. The following detailed description, in conjunction with specific accompanying drawings, illustrates the embodiments.

[0036] In order to accurately define the technical scope of this application, the core terms are defined and explained as follows:

[0037] System Memory: In this invention, system memory is a self-evolving object. It is a collection of structured knowledge assets continuously accumulated and updated by multiple learning agents, including: an indicator template library (validated query logic reuse templates), knowledge update records (corrections and supplements to the ONN model layer), a user profile library (personalized understanding configurations for each user), and a hot data cache (a fast response channel for high-frequency queries). System memory is not private data of any particular agent, but rather a public resource that can be consumed and used by all agents.

[0038] Metric Templates: Reusable query logic units extracted from historical queries after passing double-blind quality control verification. A metric template contains a complete OQS (Online Query Specification) structure and a corresponding pipelined DSL (Domain-Specific Language). Metric templates are divided into two categories: fixed metric templates (all parameters of the query question are fixed, directly reusing the complete DSL) and dynamic metric templates (parameters such as object classes, filtering conditions, and extraction fields in the query question are variable, achieving reuse through parameterized placeholders).

[0039] Coverage Score: A quantitative metric that measures the semantic matching degree between a user query and an existing metric template. The coverage score is calculated using a semantic alignment algorithm. When the score exceeds a preset threshold, the system directly retrieves the matching metric template from the hot data cache, without needing to re-execute the ABC parsing and DSL generation process.

[0040] Three-Way Parallel Dispatch: Upon receiving a verified user query, the system simultaneously initiates three processing paths: fixed-indicator hot data retrieval (exact matching), dynamic-indicator hot data retrieval (parameterized matching + semantic alignment), and ABC deep parsing. Based on the coverage score of each path, the system selects the optimal result to return, achieving intelligent scheduling of "instant response if hot data is hit, deep parsing if no match is found."

[0041] Knowledge Backward Update: When feedback signals such as quality inspection failure, manual error correction, or business rule changes occur during system operation, the knowledge learning agent systematically analyzes and classifies these feedback signals and updates them back to the ONN model layer (including correcting CBC constraints, supplementing synonym mappings, adding default rules, and improving business knowledge descriptions), so as to continuously refine the ontology semantic network.

[0042] User Profile: A personalized understanding configuration automatically built by the system based on a user's query history. The user profile consists of three dimensions: the user's frequently used ontology categories (which object classes the user typically queries), the user's organization and job description (determining their permissions and perspective), and the user's explicitly requested preferences and habits from past conversations. During the intent clarification phase, the user profile provides personalized contextual supplementation to the AI ​​agent, enabling the system to more accurately understand the user's query intent.

[0043] The QA-Driven Self-Evolution Loop is the core closed-loop mechanism of this invention. Double-blind QA generates two types of signals: query results that pass (high score) trigger the indicator learning agent to be stored as indicator templates; query results that fail trigger the knowledge learning agent to analyze the reasons for errors and update the ONN in reverse. These two types of signals drive the system's "efficiency optimization" and "accuracy correction" respectively, forming a positive cycle.

[0044] Ontology Semantic Data Network (ONN): A computable semantic network data structure with ontologies as nodes, semantic relations as edges, and CBC (Context-Behavior-Constraint) as the unified logical kernel. This network expresses relationships between business objects through graph-structured semantic connections and reasoning. The term "network" specifically refers to the semantic graph topology composed of ontology classes, relation classes, instance nodes, and their attributes, rather than a deep learning parameterized model based on artificial neuron weights and backpropagation algorithms. ONN natively supports unified modeling of objects, relations, behaviors, and contexts, as well as multimodal data mounting.

[0045] The ABC paradigm (Acquire-Build-Calculate) is a methodology for breaking down complex natural language queries into standardized operational steps. Step A (Acquire Objects) identifies the object classes involved in the problem and constructs a subgraph of object relationships; Step B (Build Dataset) determines the filtering criteria and data fields to be extracted from the objects in the subgraph; and Step C (Calculate Metrics) determines the calculation formulas between the fields.

[0046] ONN Query Specification (OQS): A structured intermediate representation output by the problem analysis agent group. It describes the query intent based on the ONN model layer in a node-edge separation format and carries all semantic information of the ABC steps.

[0047] Pipeline DSL (Domain Specific Language): A set of executable instructions for a low-level engine, which is translated from OQS by a DSL-generated agent. It uses JSON format and supports multi-step chained computation.

[0048] AI-generated structured artifacts (AI-generated structured artifacts, often referred to as "target artifacts") are machine-executable products with interpretable logical structures automatically generated by large language models or intelligent agent systems based on large language models, based on the user's natural language task description. These target artifacts include, but are not limited to: data query instructions (such as SQL, DSL, GraphQL, etc.), program code (such as Python functions, Java methods, etc.), API call sequences and their parameters, intelligent agent workflow orchestration definitions, ETL data pipeline configurations, automated test scripts, system configuration files, etc. Their common characteristic is that they possess clearly defined logical elements (operation objects, execution conditions, expected output, processing flow, etc.), and can be interpreted and translated into natural language item by item.

[0049] Reverse Semantic Restoration: A process of interpreting and translating the logical elements of a target product into natural language descriptions. Specifically, it involves restoring the operational objects in the target product into business entity names, the execution conditions into natural language constraints, the expected output into information requirement descriptions, and the processing flow into step-by-step task descriptions. These are ultimately combined into a complete "system understanding description" that describes "what the product is actually doing."

[0050] Double-Blind Quality Inspection: A verification mechanism based on information isolation. In this invention, "double-blind" specifically refers to two layers of independence guarantees: First, the quality inspection verification agent does not receive the user's original natural language task description as input when performing quality inspection, but only performs reverse semantic reconstruction based on the target product itself; Second, the quality inspection verification agent performing reverse semantic reconstruction and the generation agent performing forward generation of the target product are independent of each other, and do not share intermediate reasoning processes, prompt word content, and contextual memory. This double isolation design ensures that the quality inspection results are not affected by the anchoring effect of the original task description and the path dependency of the forward generation logic.

[0051] Semantic Consistency Comparison: This is the process of performing multi-dimensional semantic entailment analysis on the "system understanding description" generated by the quality inspection verification agent through reverse semantic reconstruction and the user's original natural language task description, in order to determine whether the two are equivalent in terms of business semantics.

[0052] Semantic Consistency Score: A comprehensive score calculated based on multi-dimensional comparison results and according to preset dimensional weights, used to quantify the degree of semantic matching between the target product and the user's original intent.

[0053] Targeted Correction: When quality inspection fails, based on the specific differences in each dimension in the multi-dimensional comparison, a feedback instruction containing a clear deviation location and correction direction is sent to the upstream generation system, so that the upstream system only corrects the parts with deviations, rather than regenerating the entire dataset.

[0054] Node-Edge Separated Structure: A graph structure representation method that describes the node information (object class, filtering conditions, extraction fields) and edge information (relationship type, direction) of the graph using independent data structures, instead of using a two-dimensional matrix representation of the adjacency matrix. This structure has a storage space complexity of O(n+e) (where n is the number of nodes and e is the number of edges), significantly reducing the token consumption required for processing large language models compared to the O(n²) of the adjacency matrix.

[0055] Condition Source Annotation (CSA): A mechanism that records both the condition value and its source in the filtering criteria. The source of each filtering condition is annotated as either "User Question" (explicitly raised by the user) or "Business Default Rule" (automatically injected by CBC constraints in the ontology model layer), making the ontology-driven condition injection process fully traceable and auditable.

[0056] Ontology Semantic Data Network (ONN) is a computable semantic network data structure with ontology classes as nodes, semantic relation classes as edges, and a CBC (Content Contextualization) pattern as its unified logical kernel. ONN consists of three layers: an ontology model layer (defining object classes, relation classes, and their CBC strategies), an instance network layer (storing specific instance data and automatically inheriting the CBC strategies from the model layer), and a behavior implementation layer (associating semantic behaviors with specific execution implementations). The term "network" specifically refers to the semantic graph topology structure composed of ontology classes, relation classes, instance nodes, and their attributes and strategies, rather than a deep learning parameterized model based on artificial neuron weights and backpropagation algorithms.

[0057] Ontology Class (OC): An abstract type definition for business objects in the ONN model layer; it serves as a "template" or "blueprint" for things. Each ontology class contains: a class identifier, a classpath, a set of attribute definitions, a semantic boundary description, and a set of associated CBC strategies. Ontology classes are divided into entity classes (Nouns, describing static objects such as equipment and personnel) and event classes (Verbs, describing dynamic activities such as process execution and quality inspection).

[0058] Link Class (LC): An abstract type definition of business relationships in the ONN model layer. Each link class contains: a source ontology class, a target ontology class, a relationship cardinality (one-to-one / one-to-many / many-to-many), relationship constraints, and optional attribute definitions. A link class is not just a simple "connection"; it can also attach independent attributes and CBC strategies, becoming a computable semantic entity.

[0059] CBC (Context-Behavior-Constraint) pattern: The unified logical core of ONN, endowing each ontology class and relation class with context awareness, behavior execution, and constraint governance capabilities. Its three dimensions are: Context defines the environmental boundaries and preconditions for behavior and constraint to take effect; Behavior defines the callable operations that encapsulate business logic, described in natural language for easy understanding by AI agents; and Constraint defines the business rule checks that must be passed before executing a behavior.

[0060] Policy (Pol): The carrier unit of CBC patterns. A policy contains a complete set of three-dimensional CBC definitions (contextual conditions, behavioral descriptions, and constraint rules), which can be associated with one or more ontology classes or relation classes. The contextual, behavioral, and constraint descriptions in the policy are in natural language, which AI agents can directly read and understand without reverse engineering the program code.

[0061] Behavior Implementation Layer: This layer acts as an "execution bridge" connecting the ONN semantic world with the enterprise's existing IT and AI capabilities. Each behavior in the strategy is described in natural language at the semantic layer (for easy AI understanding) and can be associated with one or more specific implementations. Implementations are divided into two categories: traditional IT implementations (program code functions: executed directly within the platform; API calls: driving external microservices or enterprise applications such as ERP and MES) and AI-native implementations (Agent workflows: triggering AI agents to complete complex tasks; MCP services: connecting external tools and data sources through model context protocols; A2A collaboration: coordinating execution with other AI agents through inter-agent communication protocols; Skills packages: predefined reusable AI capability modules).

[0062] Instance Inheritance: The core mechanism in the ONN instance network layer. Each ontology instance (OI) automatically inherits all CBC strategies defined in the model layer of its parent ontology class, eliminating the need for redefinition. When the CBC strategy in the model layer changes, the change is automatically propagated to all corresponding instances.

[0063] AI-Native Read-Write Loop: ONN's unique AI interaction mode. The AI ​​agent can achieve a complete "read-understand-write" closed loop in ONN: read CBC natural language descriptions to understand business rules, make inferences and decisions based on understanding, and securely invoke rules-protected behaviors to execute business operations.

[0064] exist Figure 1 In this application, an embodiment provides a multi-agent cooperative self-evolution method based on ontology semantic networks, including the following steps:

[0065] The indicator learning module is used to extract fixed and dynamic indicator templates from high-scoring quality inspection queries and manage hot data caching.

[0066] The knowledge-backward update module receives feedback signals and updates the ONN model layer.

[0067] Utilize the user profile module to build and maintain multi-dimensional user profiles and provide personalized context when clarifying intent.

[0068] In the above technical solution, a fixed indicator template and a dynamic indicator template are extracted from high-scoring quality inspection queries by using an indicator learning module and hot data cache is managed; a knowledge reverse update module is used to receive feedback signals and update the ONN model layer; a user profile module is used to build and maintain multi-dimensional user profiles and provide personalized context when intent is clarified; a self-evolutionary closed loop driven by quality inspection feedback is realized, which enables the system to continuously learn from the success and failure of each query and automatically improve query accuracy and response efficiency.

[0069] Specifically, the beneficial effects include:

[0070] Quality control is implemented before learning begins. Indicator templates are only selected from high-scoring queries in the quality control process, ensuring that all templates memorized by the system have undergone quality verification and preventing "learning the wrong ones".

[0071] The dual-type templates offer broad coverage. Fixed indicator templates support exact match, while dynamic indicator templates support new queries with similar structures through parameterized placeholders. The combination of the two significantly improves template coverage.

[0072] Three-way parallel intelligent scheduling. Fixed matching, dynamic matching, and deep parsing are performed in parallel. The system automatically selects the optimal path based on the coverage score, balancing response speed and query accuracy.

[0073] Knowledge is updated in reverse to form a closed loop. Feedback signals such as quality inspection failures, manual corrections, and business changes are automatically transformed into precise corrections to the ONN model layer, and the system's knowledge base is continuously refined.

[0074] A secure change management mechanism. Automatic modifications to the ONN model layer follow a secure process of change logging → impact analysis → gray-scale verification → rollback capability to avoid cascading effects caused by erroneous modifications.

[0075] Multi-dimensional user profiling enhances personalized understanding. A three-dimensional profile, combining ontology classification, organizational job descriptions, and historical memory preferences, enables the system to "understand users better the more it's used," significantly reducing the need for clarification and follow-up questions.

[0076] The query logic is cached, not the data results. Hot data is cached using the OQS+DSL template, not the query results. The DSL is re-executed to retrieve the latest results on each call, ensuring data real-time performance while maintaining caching efficiency.

[0077] System memory serves as a public knowledge asset. Self-evolving outputs (templates, knowledge, profiles, hot data) are shared memory at the system level, which can be consumed and used by all agents, rather than being the private data of a particular agent.

[0078] In one specific implementation scheme, it also includes:

[0079] The scheduling module is used to initiate fixed matching, dynamic matching, and deep parsing in parallel, and the optimal result is selected based on the coverage score.

[0080] In one specific implementation scheme, the indicator template includes:

[0081] Fixed indicator templates allow for direct reuse of complete OQS and DSL;

[0082] Dynamic indicator templates are reused through parameterized placeholders.

[0083] In a specific feasible implementation, knowledge reverse updates modify the ONN model layer in accordance with a safe change process that includes change logging, impact analysis, gray-scale verification, and rollback capability.

[0084] In a specific feasible implementation, the content of knowledge reverse update includes: supplementing synonym mapping, correcting CBC constraints, adding default rules, and improving business knowledge descriptions.

[0085] In a specific feasible implementation, user personas include dimensions such as the range of ontology categories commonly used by the user, the user's organizational and job descriptions, and the user's preferences for requesting the system to remember their past conversations.

[0086] In one feasible implementation, the object of hot data caching is the query logic template.

[0087] In a specific implementation plan, the high-scoring signal triggering indicator learning agent accumulates the template, while the low-scoring signal triggering knowledge learning agent updates the ONN in reverse.

[0088] In one specific implementation scheme, the multi-agent cooperative self-evolution method based on ontology semantic networks includes the following steps:

[0089] Step 1: Quality Inspection-Driven Indicator Learning – Automatic Generation of High-Frequency Query Templates

[0090] The Metric Learning Agent performs the following operations:

[0091] 1. Pre-quality inspection filtering. Only query results with a double-blind quality inspection score reaching a preset threshold (e.g., ≥90 out of 100) are accepted as learning input. Queries with low quality inspection scores are excluded from template accumulation, ensuring that all indicator templates accumulated in the system's memory have undergone quality verification.

[0092] 2. Indicator Template Classification. Query patterns that pass quality inspection are analyzed and classified into two categories:

[0093] Category 1: Fixed indicator templates. All parameters of the query (object classes, filtering conditions, extraction fields, calculation methods) are completely fixed, and each execution produces results with the same structure.

[0094] Example: "Statistics on the distribution of current faculty and staff by professional title level" - The object class (faculty and staff, professional title change records), conditions (current, highest current), field (professional title level), and calculation (count + group_by) are all fixed.

[0095] The second type: Dynamic indicator templates. The core query logic structure is fixed, but the parameters involved, such as the object class, the specific object name / time range / value range in the filtering conditions, and the selection of extraction fields, are variable. The system abstracts the variable parts as parameterized placeholders.

[0096] Example: "Statistics on the publication of papers by {department}" - the query logic structure remains unchanged, but {department} is a variable parameter; "Query the batch quality inspection pass rate of raw materials supplied by {supplier} during {time period}" - {supplier} and {time period} are variable parameters.

[0097] 3. Template Standardization. The identified patterns are standardized.

[0098] Fixed indicator template: Fully saves the OQS structure and DSL code, and annotates trigger keywords and semantic feature vectors.

[0099] Dynamic indicator template: Replace the variable parts in OQS and DSL with named placeholders (e.g., ... , ), which records the type constraints and value range of each placeholder.

[0100] 4. Incorporate into hot data cache. Store the standardized indicator templates in the hot data cache layer and record the following metadata: template creation time, quality inspection score, number of times it was called, time of the most recent call, and source query text.

[0101] Step Two: Three-way Parallel Scheduling – Intelligent Matching and Utilization of Hot Data

[0102] When a new query enters the system, the scheduling agent simultaneously initiates three paths:

[0103] 1. Path 1: Fixed Metric Exact Match. The user query is semantically aligned with the trigger keywords and semantic feature vectors of the fixed metric template. If a template with a perfect semantic match is found (coverage score of 100%), the DSL of that template is directly retrieved, executed, and the latest results returned.

[0104] 2. Path Two: Dynamic Metric Parameterized Matching. The user query is semantically aligned and compared with the structural features of the dynamic metric template. If a structurally matching template is found (coverage score ≥ preset threshold), variable parameter values ​​are extracted from the user query, filled into placeholders in the template, and the complete DSL is generated and executed.

[0105] Example: The dynamic template is "Statistics" When a user asks for statistics on the publication of papers by the School of Computer Science, the system recognizes org_name="School of Computer Science", fills in the placeholders, and then executes the DSL.

[0106] 3. Path Three: ABC Deep Parsing. When neither Path One nor Path Two finds a matching template (or the coverage score is below the threshold), the complete ABC paradigm parsing process is initiated—four dedicated agents work together serially to generate the OQS, which is then translated into DSL.

[0107] 4. Optimal Path Selection. The scheduling agent selects the optimal path based on its coverage score: fixed matching with 100% coverage is prioritized, followed by dynamic matching with high coverage, and finally ABC deep parsing. As the coverage of the hot data cache expands with system usage time, the proportion of hot data hits continuously increases, and the overall system response speed continuously improves.

[0108] Step 3: Knowledge Reverse Update ONN – Refining Ontology Semantics from Feedback

[0109] The knowledge learning agent receives the following four types of feedback signals, systematically analyzes them, and updates the ONN model layer in reverse:

[0110] (a) Case analysis of quality inspection failures.

[0111] When double-blind quality inspection identifies semantic bias, the knowledge learning agent analyzes the causes of the bias and categorizes them for processing:

[0112] Object class identification error → Supplement the semantic boundary description or synonym mapping of the object class in the ONN model layer.

[0113] Example: The user says "mentor" but the system fails to map it to the "faculty and staff" class → supplement the synonym mapping "mentor → faculty and staff".

[0114] Missing or incorrect conditions → Correct the default rules in the CBC constraints of this object class.

[0115] Example: The system has not added the default condition "last_flag=Yes" to "Professional Title Change Record" → Add this business default rule to the CBC constraint.

[0116] Relationship path error → Check and correct the relation class definitions in the ONN model layer or add missing relations.

[0117] (b) Manual error correction feedback.

[0118] When a user or administrator manually corrects the query results, the knowledge learning agent records the corrections, analyzes the reasons for the corrections, and updates the ONN according to the same classification mechanism.

[0119] (c) Changes to business rules.

[0120] When business rules change (such as adding a new type of talent honor level or modifying the judgment criteria for employment status), the Knowledge Learning Agent receives the change notification and automatically updates the CBC constraints and attribute definitions of the corresponding object classes in the ONN model layer.

[0121] (d) The description of business knowledge is comprehensive.

[0122] In addition to constraints, synonyms, and rules, the knowledge learning agent is also responsible for refining the business knowledge descriptions in the ONN model layer—including semantic boundary descriptions of object classes, explanations of the business meanings of attributes, and interpretations of the business context of relationships. This natural language-described knowledge continuously enriches the semantic layer of ONN, enabling the AI ​​agent to gain an increasingly deeper understanding of the business domain.

[0123] (e) Safety changes and rollbacks.

[0124] Any automatic modifications to the ONN model layer follow the following safety mechanisms: Change logging – recording the time, reason, and values ​​before and after each modification; Impact analysis – assessing the query range that the modification may affect; Gray-scale verification – performing regression tests on historical queries within the affected range after modification; Rollback capability – if regression testing finds that the modification causes a decrease in the accuracy of other queries, automatically rolling back to the version before the modification.

[0125] Step 4: User Profile Building and Intent Personalization

[0126] The user profiling analysis agent automatically builds and maintains a personalized profile for each user, which consists of the following three dimensions:

[0127] (a) Ontology classification scope. Based on the user's historical query records, the most frequently queried ontology classes (object classes) and their distribution are statistically analyzed.

[0128] Example: The ontology classification scope for users in the Research Office is {papers: 60%, research projects: 25%, faculty and staff: 15%}; the scope for users in the Academic Affairs Office is {courses: 40%, grades: 35%, students: 25%}.

[0129] Function: When a user submits a fuzzy query, the system prioritizes object class identification within the user's commonly used categories to reduce ambiguity.

[0130] (b) Organization and Job Description. Record the user's organizational level, department, job role, and corresponding data permissions and perspectives.

[0131] Example: User "Professor Zhang", Organization = School of Computer Science, Position = Professor / Master's Supervisor, Focus = Students under his supervision + His own papers + Projects.

[0132] Function: When Professor Zhang asks "How many papers have my students published?", the system automatically interprets "my" as "Master's students supervised by Professor Zhang", without needing to ask further.

[0133] (c) Historical dialogue memory requirements. User preferences and habits regarding which the system explicitly requests the system to "remember" historical dialogues.

[0134] Example: A user says, "When compiling statistics on papers in the future, I will only look at those from CSCD and above by default." → The system records this preference in the user profile and applies it automatically in subsequent queries.

[0135] Purpose: To reduce repetitive condition descriptions, enabling the system to "understand the user better with use." Application in the intent clarification phase: When receiving a user query, the intent clarification agent simultaneously reads the user's profile as supplementary context.

[0136] Image information helps AI agents: narrow the scope of object class recognition (reduce ambiguity);

[0137] Automatically complete implicit constraints (such as organization scope and permission filtering);

[0138] This incorporates users' personalized preferences (such as default level filtering). The ultimate goal is to reduce unnecessary rounds of clarification and follow-up questions, thereby improving the user experience.

[0139] Step 5: Hot Data Cache Management and Eviction Strategy

[0140] (a) Cache Objects. The hot data cache stores the query logic template (OQS+DSL), not the query result data. This design allows the system to obtain the latest results by re-executing the cached DSL even if the underlying data changes, balancing caching efficiency and data real-time performance.

[0141] (b) Caching hierarchy. Divided into two levels according to matching method:

[0142] Fixed metric caching – when a DSL is hit, it executes directly without any parameterization, resulting in the fastest response time.

[0143] Dynamic metric caching requires extracting parameters and filling placeholders before execution upon hit, resulting in a slower response time.

[0144] (c) Hybrid eviction policy. A hybrid eviction policy combining LRU (Least Recently Used) and call frequency is adopted:

[0145] Prioritize retaining high-frequency, high-scoring templates (high number of calls + high quality inspection score);

[0146] Eliminate templates that have not been used for a long time (templates that have not been invoked within a set time window);

[0147] Regularly conduct quality checks and re-inspections on existing templates (business rules may change, causing old templates to become invalid), and remove those that fail.

[0148] Step Six: Forming a complete self-evolutionary closed loop

[0149] The self-evolutionary closed loop of this invention consists of the following four positive loops:

[0150] Loop 1 (Efficiency Loop): Query → High-scoring quality inspection → Template accumulation → Hot data caching → Next hit → Instant response → Efficiency improvement.

[0151] Cycle Two (Accuracy Cycle): Query → Low Quality Inspection Score → Error Analysis → Knowledge Update ONN → More Accurate Next Time → Improved Accuracy.

[0152] Cycle 3 (Understanding Cycle): Query → User Profile Accumulation → Personalized Intent Understanding → Reduce Follow-up Questions → Improved Experience.

[0153] Cycle Four (Knowledge Base Cycle): Manual error correction / business change → Knowledge learning → ONN refinement → Deepening of overall semantic understanding. The combined effect of the four cycles: The longer the system is used → the higher the coverage of hot data, the more accurate the ONN knowledge, and the richer the user profile → the system becomes "faster, more accurate, and more user-savvy the more it is used."

[0154] In a specific feasible implementation plan, and in conjunction with an example of indicator template accumulation for a specific university scientific research management scenario, the technical solution of the present invention will be described in detail. This includes:

[0155] Scenario: The process of accumulating indicator learning agents after three months of operation of the university's scientific research management system.

[0156] Example of a fixed indicator template:

[0157] Frequently asked question by users: "Statistics on the distribution of talent honors and achievements among faculty and staff of various colleges."

[0158] → ABC parsing generates OQS (3 object classes: faculty and staff, organization, talent honors and achievements; 3 relationships; calculate metrics including count + group_by).

[0159] → Double-blind quality inspection score of 95 points (≥90 points threshold)

[0160] → Metric Learning Agent Determination: All parameters are fixed (no user-inputted variables required)

[0161] → Set as a fixed indicator template: template_id=FIX-001, keywords=[“Talent Honors”,“College”,“Distribution”]

[0162] → Include in hot data cache → The next time a similar question is asked, path one will be an exact match, and the DSL will be executed directly to return the latest data. Example of dynamic indicator template accumulation:

[0163] User Question 1: "The number of SCI papers published by the School of Statistics and Computer Science in 2024?"

[0164] User Question 2: "The number of SCI papers published by the School of Statistics and Mathematics in 2024?"

[0165] → The two queries have the same structure, only the "department name" is different.

[0166] → Both quality inspection scores are ≥90 points

[0167] → After the indicator learns the Agent's recognition pattern, it is abstracted into a dynamic template:

[0168] {

[0169] "template_id": "DYN-001",

[0170] "pattern": "statistics" Published in 1998 Number of papers

[0171] "params": [

[0172] {"name": "org_name", "type": "organization.name", "required": true},

[0173] {"name": "year", "type": "integer", "required": true},

[0174] {"name": "journal_level", "type": "paper.level", "required": false, "default": null}

[0175] ],

[0176] "oqs_template": {

[0177] "nodes": [

[0178] {"id": "n1", "class": "paper", "filters": [

[0179] {"field": "pub_year", "op": "=", "value": ,

[0180] {"field": "level", "op": "=", "value": "if_present": true}

[0181] ], "extract_fields": ["title"]},

[0182] {"id": "n2", "class": "faculty and staff", "filters": [{"field": "status", "op":"=", "value": "on duty", "source": "default business rule"}]},

[0183] {"id": "n3", "class": "Organization", "filters": [{"field": "name", "op":"=", "value":

[0184] ],

[0185] "edges": [

[0186] {"from": "n2", "to": "n1", "type": "author_of"},

[0187] {"from": "n2", "to": "n3", "type": "belongs_to"}

[0188] ],

[0189] "metrics": {"expr": "COUNT(n1)", "group_by": []}

[0190] }

[0191] }

[0192] → When a user asks "the number of papers published by the School of Physics in 2025", the second path matches the template DYN-001, fills in org_name="School of Physics" and year=2025, and directly generates the DSL for execution.

[0193] 8.2 Example 2: Knowledge Reverse Update in Factory Quality Traceability Scenario

[0194] Scenario: During the operation of ONN in an engine manufacturing plant, the knowledge learning agent processes feedback.

[0195] Cases of quality inspection failure:

[0196] User question: "I need to check the quality inspection results for all crankshaft batches processed by the CNC machine tool on the 5th of last month."

[0197] → During ABC parsing, the system failed to correctly identify "CNC Machine Tool No. 5" as an instance of the Equipment object class (mistakenly interpreting "No. 5" as a batch number prefix).

[0198] → Double-blind quality control score: 55 points (below the threshold)

[0199] → Knowledge Learning Agent Analysis: Error Type = Object Class Identification Error

[0200] → Reverse update ONN: Add "Numbering pattern description: Equipment numbers are usually in the format '{N}{{equipment type}', such as 'CNC machine tool No. 5'" to the semantic boundary description of the Equipment object class.

[0201] → Also add a synonym mapping: "CNC machine tool → Numerical control machine tool → Machining center"

[0202] → Change Verification: Regression testing was performed on historical queries containing device numbers → Accuracy improved from 68% to 95%; Business knowledge description improved:

[0203] During operation, it was found that the AI ​​agent did not have a deep enough understanding of the business implications of "key components".

[0204] → The Knowledge Learning Agent supplements the business knowledge description in the semantic boundary of the Part object class: "Critical components (critical_flag=true) refer to components that have a direct impact on the safety and reliability of the product. The quality inspection of critical components must be carried out with 100% full inspection (not sampling inspection), and the inspection personnel must hold critical component inspection qualifications."

[0205] → This business knowledge description is stored in the ONN model layer in natural language, which the AI ​​agent can directly read and understand when processing key component-related queries. 8.3 Example 3: User Profile Optimization Intent Understanding

[0206] Scenario: The construction and application of Professor Zhang's user profile in a university research management system.

[0207] Portrait creation process:

[0208] After 50 queries, the user profile analysis agent constructed the following profile for Professor Zhang:

[0209] {

[0210] "user_id": "PROF-ZHANG-001",

[0211] "ontology_scope": {

[0212] "Common Categories": ["Thesis: 45%", "Master's Students: 30%", "Research Projects: 20%", "Faculty and Staff: 5%"],

[0213] "Never Queryed": ["Organization", "Course", "Grade"]

[0214] },

[0215] "organization": {

[0216] "department": "School of Computer Science"

[0217] "position": "Professor / Master's Supervisor",

[0218] "data_scope": "Students I supervised + My own papers and projects"

[0219] },

[0220] "memorized_preferences": [

[0221] {"Time": "2025-03-01", "Content": "Paper statistics only consider CSCD and above levels by default"}

[0222] {"Time": "2025-04-15", "Content": "Student search includes graduated students"} ]

[0224] }

[0225] Comparison of portrait application effects:

[0226] Professor Zhang asked, "What papers have my students published recently?" Without a profile: The system needs to follow up with questions such as, "Does 'my students' refer to the master's students you supervise?", "How long ago is 'recently'?", and "What level of paper is required?"—at least three rounds of follow-up questions.

[0227] When a profile is available: The system reads the profile → automatically understands "My Students" = Master's students supervised by Professor Zhang (including those who have graduated), "Recent" = the most recent year (system default), thesis level = CSCD and above (user preference) → directly returns the result, 0 rounds of follow-up questions.

[0228] The beneficial effects of the above technical solution include:

[0229] The hit rate of hot data continues to improve. As the system is used for longer periods, the coverage of indicator templates continues to expand. It is expected that after 6 months of system operation, the hit rate of hot data for high-frequency queries will reach over 60%, significantly reducing the number of calls to ABC deep analysis.

[0230] ONN knowledge accuracy continues to improve. The knowledge learning agent is driven by continuous feedback to update, and the synonym mapping, CBC constraints, and business knowledge descriptions in the ONN model layer are continuously refined, resulting in continuous improvement in query accuracy over time.

[0231] User experience has been continuously optimized. User profiling has reduced the average number of clarification and follow-up questions from 3.2 rounds to 0.8 rounds, making users feel that the system "understands them more and more".

[0232] Computational resource savings. When hot data is hit, the entire process of ABC parsing and DSL generation is skipped, reducing the computational resource consumption of a single query to about 5% of that of deep parsing.

[0233] exist Figure 2 In this application, embodiments provide a multi-agent cooperative self-evolutionary system based on an ontology semantic network, comprising:

[0234] The indicator learning module is used to extract fixed indicator templates and dynamic indicator templates from high-scoring quality inspection queries and manage hot data cache;

[0235] The knowledge reverse update module is used to receive feedback signals and update the ONN model layer;

[0236] The user profile module is used to build and maintain multi-dimensional user profiles and provide personalized context when intent is clarified.

[0237] In the above technical solution, a fixed indicator template and a dynamic indicator template are extracted from high-scoring quality inspection queries by using an indicator learning module and hot data cache is managed; a knowledge reverse update module is used to receive feedback signals and update the ONN model layer; a user profile module is used to build and maintain multi-dimensional user profiles and provide personalized context when intent is clarified; a self-evolutionary closed loop driven by quality inspection feedback is realized, which enables the system to continuously learn from the success and failure of each query and automatically improve query accuracy and response efficiency.

[0238] In one specific implementation scheme, it also includes:

[0239] The scheduling module is used to initiate fixed matching, dynamic matching, and deep parsing in parallel and select the optimal result based on the coverage score.

[0240] Those skilled in the art will know that this application can be implemented as a system, method, or computer program product.

[0241] Therefore, this disclosure can be implemented in the following forms: it can be entirely hardware, entirely software (including firmware, resident software, microcode, etc.), or a combination of hardware and software, generally referred to herein as a "circuit," "module," or "system." Furthermore, in some embodiments, this application can also be implemented as a computer program product contained in one or more computer-readable media, which contains computer-readable program code. Any combination of one or more computer-readable media can be used. A computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. A computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this document, a computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in connection with an instruction execution system, apparatus, or device.

[0242] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of this application. Based on this, various substitutions and improvements can be made to this application, all of which fall within the protection scope of this application.

Claims

1. A multi-agent cooperative self-evolution method based on an ontology semantic data network, wherein the ontology semantic data network is a computable semantic network data structure with ontologies as nodes, semantic relations as edges, and a CBC (Context-Behavior-Constraint) as a unified logical kernel, wherein the CBC refers to the context-behavior-constraint logic, characterized in that... Includes the following steps: The indicator learning module is used to extract fixed and dynamic indicator templates from high-scoring quality inspection queries and manage hot data caching. The indicator templates include: a fixed indicator template, which directly reuses the complete ONN Query Specification (OQS) and the corresponding pipelined DSL; wherein, the ONN Query Specification (OQS) is a structured intermediate representation output by the problem analysis agent group, which uses a node-edge separation format to describe the query intent of the ontology model layer based on the ontology semantic data network, and carries all the semantic information of the three steps of the ABC paradigm; and a dynamic indicator template, which is reused through parameterized placeholders; wherein, the ABC paradigm is a methodology for decomposing complex natural language queries into standardized operational steps, wherein step A identifies the object classes involved in the problem and constructs an object relationship subgraph; step B determines the filtering conditions and data fields to be extracted from the objects in the subgraph; and step C determines the calculation formula between the fields; the pipelined DSL is an executable instruction set of the underlying engine translated by the DSL generation agent from the OQS, which uses JSON format and supports multi-step chained computation; The knowledge reverse update module receives feedback signals and updates the ontology model layer of the ontology semantic data network. The ontology semantic data network includes: an ontology model layer, which defines object classes, relation classes and their CBC strategies; an instance network layer, which stores specific instance data and automatically inherits the CBC strategies of the model layer; and a behavior implementation layer, which associates semantic behaviors with specific execution implementations. The content of knowledge reverse update includes: supplementing synonym mapping, correcting CBC constraints, adding default rules, and improving business knowledge descriptions; among which, the correction of CBC constraints refers to the correction of the "context-behavior-constraint" logic of the ontology semantic data network; Utilize the user profile module to build and maintain multi-dimensional user profiles and provide personalized context when clarifying intent.

2. The multi-agent cooperative self-evolution method based on ontology semantic data network according to claim 1, characterized in that, Also includes: The scheduling module is used to initiate fixed matching, dynamic matching, and deep parsing in parallel, and the optimal result is selected based on the coverage score.

3. The multi-agent cooperative self-evolution method based on ontology semantic data network according to claim 2, characterized in that, Knowledge reverse updates modify the ontology model layer of the ontology semantic data network according to a safe change process that includes change logging, impact analysis, gray-scale verification, and rollback capability.

4. The multi-agent cooperative self-evolution method based on ontology semantic data network according to claim 3, characterized in that, User profiles include dimensions such as the range of ontology categories commonly used by users, the user's organizational and job descriptions, and the user's preferences for requesting the system to remember their past conversations.

5. The multi-agent cooperative self-evolution method based on ontology semantic data network according to claim 4, characterized in that, The object cached for hot data is the query logic template.

6. The multi-agent cooperative self-evolution method based on ontology semantic data network according to claim 5, characterized in that, The high-scoring signal triggering indicator learning agent accumulates templates, while the low-scoring signal triggering knowledge learning agent updates the ontology semantic data network in reverse.

7. A multi-agent cooperative self-evolutionary system based on an ontology semantic data network, wherein the ontology semantic data network is a computable semantic network data structure with ontologies as nodes, semantic relations as edges, and a CBC (Context-Behavior-Constraint) as a unified logical kernel, wherein the CBC refers to the context-behavior-constraint logic, characterized in that... include: The indicator learning module is used to extract fixed indicator templates and dynamic indicator templates from high-scoring quality inspection queries and manage hot data cache; The indicator templates include: a fixed indicator template, which directly reuses the complete ONN query specification OQS and the corresponding pipelined DSL; wherein, the ONN query specification OQS is a structured intermediate representation output by the problem analysis agent group, which uses a node-edge separation format to describe the query intent of the ontology model layer based on the ontology semantic data network, and carries all the semantic information of the three steps of the ABC paradigm; and a dynamic indicator template, which is reused through parameterized placeholders; wherein, the ABC paradigm is a methodology for decomposing complex natural language queries into standardized operation steps, wherein step A identifies the object classes involved in the problem and constructs an object relationship subgraph; step B determines the filtering conditions and data fields to be extracted from the objects in the subgraph; and step C determines the calculation formula between the fields; and the pipelined DSL is an executable instruction set of the underlying engine translated by the DSL generation agent from the OQS, which uses JSON format and supports multi-step chained computation; The knowledge reverse update module is used to receive feedback signals and update the ontology model layer of the ontology semantic data network. The ontology semantic data network includes: an ontology model layer, which defines object classes, relation classes and their CBC strategies; an instance network layer, which stores specific instance data and automatically inherits the CBC strategies of the model layer; and a behavior implementation layer, which associates semantic behaviors with specific execution implementations. The content of knowledge reverse update includes: supplementing synonym mapping, correcting CBC constraints, adding default rules, and improving business knowledge descriptions; among which, the correction of CBC constraints refers to the correction of the "context-behavior-constraint" logic of the ontology semantic data network; The user profile module is used to build and maintain multi-dimensional user profiles and provide personalized context when intent is clarified.

8. The multi-agent cooperative self-evolutionary system based on ontology semantic data network according to claim 7, characterized in that, Also includes: The scheduling module is used to initiate fixed matching, dynamic matching, and deep parsing in parallel and select the optimal result based on the coverage score.

Images